Anthropic enters the custom silicon race to optimize Claude AI, hiring specialized engineers to lower query costs while maintaining a multi-chip strategy.
Anthropic is hiring engineers to build a custom silicon team, marking its first public move into proprietary hardware design. The shift is driven by the brutal arithmetic of serving billions of tokens daily to a user base that supports a $30 billion revenue run-rate.
Job listings for silicon engineers reveal a salary range between $320,000 and $485,000. The company is specifically seeking veterans who have shipped finished semiconductor designs and can make high-stakes technical decisions without the safety net of a massive corporate organization.
This recruitment drive transforms months of industry speculation into a staffed program. While Reuters previously reported that the lab was exploring proprietary chips, the current hiring push for front-end design and pre-silicon verification experts confirms the project has moved into active development.
Hardware and software co-design
Anthropic aims to co-design its hardware and models simultaneously. This integration is intended to increase the performance and scalability of the Claude family of models, effectively lowering the cost of every individual query.
Despite this push, the company is not attempting a total exit from the existing ecosystem. A spokesperson told unite.ai that Anthropic will maintain a multi-chip approach. The lab continues to utilize massive TPU capacity from Google and Broadcom, alongside hardware from AWS, Nvidia, and AMD.
This strategy allows Anthropic to match specific workloads to the most efficient chip available. By diversifying its hardware stack, the company avoids the risk of total dependency on a single vendor while it develops its own internal capabilities.
Industry analysts estimate that developing a cutting-edge AI chip can cost roughly $500 million before manufacturing even begins. While Anthropic has not disclosed a release date for its first chip, previous reports from TechCrunch suggested the company had held discussions with Samsung Electronics as a potential manufacturing partner.
Vertical integration as a survival trait
Anthropic joins a growing list of frontier labs, including OpenAI and Meta, that are pursuing custom silicon. For these companies, artificial intelligence is no longer just a software challenge but a power and thermal management problem. Controlling the silicon allows a lab to squeeze more tokens per watt, which is the only sustainable way to scale to hundreds of millions of users.
This move also serves as a hedge against the volatility of the semiconductor supply chain. As generative AI services expand, the demand for advanced chips often outstrips supply, leaving labs at the mercy of Nvidia's allocation schedules.
The broader competitive landscape remains aggressive. CNBC reports that OpenAI is similarly orienting toward high-productivity enterprise use cases to capture market share from rivals like Google and Anthropic. In this environment, the ability to lower the cost of compute becomes a primary competitive advantage in pricing and margin.
Whether Anthropic can successfully transition from a model-centric lab to a hardware-capable entity remains to be seen. The leap from designing a neural network to shipping a physical piece of silicon is a notorious engineering hurdle that has humbled many tech giants.
FAQ
Why is Anthropic designing its own chips if it uses Nvidia?
Custom silicon allows for hardware-software co-design, which optimizes efficiency and lowers the cost per query compared to general-purpose GPUs.
Will Anthropic stop using Google and AWS hardware?
No. The company is pursuing a multi-chip strategy to ensure it can match different workloads to the most suitable processors.
How much does it cost to develop an AI chip?
Industry estimates suggest initial development costs can reach approximately $500 million before manufacturing expenses are factored in.
What is the goal of the new silicon team?
To improve the performance, scalability, and cost-efficiency of the Claude AI models at the scale required by their enterprise customers.
